freedreno/a6xx: skip tiles with no geometry
authorRob Clark <robdclark@chromium.org>
Sat, 27 Jul 2019 16:00:37 +0000 (09:00 -0700)
committerRob Clark <robdclark@chromium.org>
Fri, 2 Aug 2019 17:24:14 +0000 (10:24 -0700)
If no clear, and no geometry according to VSC_STATE[pipe] we can skip
the tile entirely.  If there is a fast-clear, we can't skip restore
(clear) or resolve IBs, but we can still skip draw IB.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
src/gallium/drivers/freedreno/a6xx/fd6_gmem.c
src/gallium/drivers/freedreno/freedreno_context.h
src/gallium/drivers/freedreno/freedreno_gmem.c

index 57f499bbb1885502cad395aeebff59e2438561c0..7ac1bae13ee4046a4145ea84ee9d4d6e05faf634 100644 (file)
@@ -545,6 +545,45 @@ check_vsc_overflow(struct fd_context *ctx)
        }
 }
 
+/*
+ * Emit conditional CP_INDIRECT_BRANCH based on VSC_STATE[p], ie. the IB
+ * is skipped for tiles that have no visible geometry.
+ */
+static void
+emit_conditional_ib(struct fd_batch *batch, struct fd_tile *tile,
+               struct fd_ringbuffer *target)
+{
+       struct fd_ringbuffer *ring = batch->gmem;
+
+       if (target->cur == target->start)
+               return;
+
+       emit_marker6(ring, 6);
+
+       unsigned count = fd_ringbuffer_cmd_count(target);
+
+       BEGIN_RING(ring, 5 + 4 * count);  /* ensure conditional doesn't get split */
+
+       OUT_PKT7(ring, CP_REG_TEST, 1);
+       OUT_RING(ring, A6XX_CP_REG_TEST_0_REG(REG_A6XX_VSC_STATE_REG(tile->p)) |
+                       A6XX_CP_REG_TEST_0_BIT(tile->n) |
+                       A6XX_CP_REG_TEST_0_UNK25);
+
+       OUT_PKT7(ring, CP_COND_REG_EXEC, 2);
+       OUT_RING(ring, 0x10000000);
+       OUT_RING(ring, 4 * count);  /* conditionally execute next 4*count dwords */
+
+       for (unsigned i = 0; i < count; i++) {
+               uint32_t dwords;
+               OUT_PKT7(ring, CP_INDIRECT_BUFFER, 3);
+               dwords = fd_ringbuffer_emit_reloc_ring_full(ring, target, i) / 4;
+               assert(dwords > 0);
+               OUT_RING(ring, dwords);
+       }
+
+       emit_marker6(ring, 6);
+}
+
 static void
 set_scissor(struct fd_ringbuffer *ring, uint32_t x1, uint32_t y1, uint32_t x2, uint32_t y2)
 {
@@ -1211,7 +1250,11 @@ fd6_emit_tile_mem2gmem(struct fd_batch *batch, struct fd_tile *tile)
 static void
 fd6_emit_tile_renderprep(struct fd_batch *batch, struct fd_tile *tile)
 {
-       fd6_emit_ib(batch->gmem, batch->tile_setup);
+       if (batch->fast_cleared || !use_hw_binning(batch)) {
+               fd6_emit_ib(batch->gmem, batch->tile_setup);
+       } else {
+               emit_conditional_ib(batch, tile, batch->tile_setup);
+       }
 }
 
 static void
@@ -1294,6 +1337,16 @@ prepare_tile_fini_ib(struct fd_batch *batch)
        }
 }
 
+static void
+fd6_emit_tile(struct fd_batch *batch, struct fd_tile *tile)
+{
+       if (!use_hw_binning(batch)) {
+               fd6_emit_ib(batch->gmem, batch->draw);
+       } else {
+               emit_conditional_ib(batch, tile, batch->draw);
+       }
+}
+
 static void
 fd6_emit_tile_gmem2mem(struct fd_batch *batch, struct fd_tile *tile)
 {
@@ -1334,7 +1387,11 @@ fd6_emit_tile_gmem2mem(struct fd_batch *batch, struct fd_tile *tile)
        OUT_RING(ring, A6XX_CP_SET_MARKER_0_MODE(RM6_RESOLVE) | 0x10);
        emit_marker6(ring, 7);
 
-       fd6_emit_ib(ring, batch->tile_fini);
+       if (batch->fast_cleared || !use_hw_binning(batch)) {
+               fd6_emit_ib(batch->gmem, batch->tile_fini);
+       } else {
+               emit_conditional_ib(batch, tile, batch->tile_fini);
+       }
 
        OUT_PKT7(ring, CP_SET_MARKER, 1);
        OUT_RING(ring, A6XX_CP_SET_MARKER_0_MODE(0x7));
@@ -1431,6 +1488,7 @@ fd6_gmem_init(struct pipe_context *pctx)
        ctx->emit_tile_prep = fd6_emit_tile_prep;
        ctx->emit_tile_mem2gmem = fd6_emit_tile_mem2gmem;
        ctx->emit_tile_renderprep = fd6_emit_tile_renderprep;
+       ctx->emit_tile = fd6_emit_tile;
        ctx->emit_tile_gmem2mem = fd6_emit_tile_gmem2mem;
        ctx->emit_tile_fini = fd6_emit_tile_fini;
        ctx->emit_sysmem_prep = fd6_emit_sysmem_prep;
index 1fcebded21b214cc6fddad7216733d1ae4740e00..43c73ef909a877b6963aa87c568061f55d47ab79 100644 (file)
@@ -310,6 +310,7 @@ struct fd_context {
        void (*emit_tile_prep)(struct fd_batch *batch, struct fd_tile *tile);
        void (*emit_tile_mem2gmem)(struct fd_batch *batch, struct fd_tile *tile);
        void (*emit_tile_renderprep)(struct fd_batch *batch, struct fd_tile *tile);
+       void (*emit_tile)(struct fd_batch *batch, struct fd_tile *tile);
        void (*emit_tile_gmem2mem)(struct fd_batch *batch, struct fd_tile *tile);
        void (*emit_tile_fini)(struct fd_batch *batch);   /* optional */
 
index 2f939b34ca9dd60cbd86724aa6f891d0fc82e96d..b1489f2e9136ca97edd68ab7d1bb9be1c0d17674 100644 (file)
@@ -376,7 +376,11 @@ render_tiles(struct fd_batch *batch)
                        ctx->query_prepare_tile(batch, i, batch->gmem);
 
                /* emit IB to drawcmds: */
-               ctx->emit_ib(batch->gmem, batch->draw);
+               if (ctx->emit_tile) {
+                       ctx->emit_tile(batch, tile);
+               } else {
+                       ctx->emit_ib(batch->gmem, batch->draw);
+               }
                fd_reset_wfi(batch);
 
                /* emit gmem2mem to transfer tile back to system memory: */